| With the rapid development of sensor technology and artificial intelligence,three-dimensional point clouds gradually occupy an increasingly important position in the field of computer vision.Threedimensional point cloud research is mainly divided into four categories:point cloud classification,point cloud segmentation,point cloud semantic segmentation,and point cloud instance segmentation.Among them,point cloud semantic segmentation is the basis for obtaining point cloud semantic information.Therefore,efficient and accurate semantic segmentation of 3D point cloud scenes is the key to current3 D scene understanding and environmental intelligence perception,which has become an increasingly important topic.The existing point cloud semantic segmentation algorithms can be divided into two categories,namely traditional methods and neural network methods.However,both types of methods have their limitations.The traditional method calculates the manual features according to manually designed operators,and then applies the manual features to the machine learning classifier to obtain the classification results.However,although the training speed of traditional methods is very fast,the further development of such methods is limited due to the complexity of feature engineering of 3D point clouds.The neural network method improves the generalization ability of the network model through a large amount of training,and is suitable for most scenarios.Nevertheless,due to their loss of detail,existing networks lack the ability to recognize complex scenes.RGAM is divided into four modules.First,the independent feature learning module is used to extract the independent features of each point.Secondly,through the neighborhood feature learning module,multiple neighborhoods in the point cloud are determined,and local information and global information are merged as domain features.Third,the feature learning module maps the neighborhood features obtained in the second step to each point,and integrates independent features as the discriminative feature of each point.Finally,the channel attention module uses the attention mechanism of the channel domain,learns the correlation between each channel of the discriminative feature and the key information,and increases the RGAM's ability to recognize complex scenes.This paper proposes a novel network architecture called the ring grouping neural network with attention module(RGAM).RGAM presents four improvements over existing networks:(1)For the uneven density of the point cloud,novel multi-scale ring grouping learning is designed to extract the multi-scale neighborhood features without overlapping sampling,allowing the network to adapt to objects of different scales.And the robustness of the network is also enhanced.(2)For the problem that a point belongs to multiple neighborhoods,neighborhood information fusion is defined by a weighted sum of multiple neighborhood features,enabling the representation of each point to be considered in different neighborhoods.In this way,the features of multiple neighborhoods are fused into each point in the neighborhood.(3)In order to use fixed relative position information of different types of objects,in the global view,a spatial attention module is introduced among the neighborhoods,allowing long-range contextual information to be exploited for 3D point cloud semantic segmentation.(4)Aiming at the problem that different types of objects with similar colors are difficult to distinguish,channel attention module is appended to RGAM: the correlation of each channel with key information enhances the complex scene recognition ability of RGAM.In this paper,the proposed network is tested on three challenging data sets of S3 DIS,Scan Net and NYU-V2,and the experimental results of this algorithm are compared with existing ones based on several latest algorithms for semantic segmentation of 3D point clouds.Network for comparison.Experimental results show that RGAM has a higher accuracy rate than several other algorithms. |
PDF Full Text Request